Wearable Sensor Feature Fusion for Human Activity Recognition (HAR): A Proposed Classification Framework

Abstract

Human Activity Recognition (HAR) focuses on detecting people's daily regular activities based on time-series recordings of their actions or motions. Due to the extensive feature engineering and human feature extraction required by traditional machine learning algorithms, they are timeconsuming to develop. To identify complicated human behaviors, deep learning approaches are more suited since they can automatically learn the features from the data. In this paper, a feature-fusion concept on handcrafted features and deep learning features is proposed to increase the recognition accuracy of diverse human physical activities using wearable sensors. The deep learning model Long-Short Term Memory based Deep Recurrent Neural Network (LSTM-DRNN) will be used to extract deep features. By fusing the handcrafted produced features with the automatically extracted deep features through the use of deep learning, the performance of the HAR model can be improved, which will result in a greater level of accuracy in the HAR model. Experiments conducted on two publicly available datasets show that the proposed feature fusion achieves a high level of classification accuracy.